Struct TripletRecipe 

pub struct TripletRecipe {
    pub name: Cow<'static, str>,
    pub anchor: Selector,
    pub positive_selector: Selector,
    pub negative_selector: Selector,
    pub negative_strategy: NegativeStrategy,
    pub weight: f32,
    pub instruction: Option<Cow<'static, str>>,
    pub allow_same_anchor_positive: bool,
}

Defines a triplet recipe (anchor/positive/negative selection + weighting).

Split-isolation contract

All three chunk slots (anchor, positive, negative) must resolve to records whose IDs hash to the same split as the request split. The sampler enforces this automatically for Selector::Role, Selector::Paragraph, and Selector::Random — those selectors always read from the record that was already confirmed to be in the correct split.

Selector::TemporalOffset crosses a record boundary (it picks a different record by proximity in time) and the split check is re-applied inside select_temporal_neighbor. No additional care is required on your side, but you should be aware that in pools with few same-split neighbors the selector will return None and fall back to skipping a slot rather than contaminating splits.

Stable IDs

Record IDs must be stable across runs. Split assignment is derived deterministically from the record ID and the sampler seed; changing an ID changes its split assignment, which invalidates any persisted split state. IDs should also be globally unique — if two records from different sources share the same ID, only one will be kept in the sampler, and the discarded record’s split assignment silently goes with it.

Fields

name: Cow<'static, str>

Unique name for this recipe.

anchor: Selector

Selector used for anchor chunks.

positive_selector: Selector

Selector used for positive chunks (same record).

negative_selector: Selector

Selector used for negative chunks (different record).

negative_strategy: NegativeStrategy

Strategy used to pick negatives.

weight: f32

Relative weight controlling how often this recipe is selected versus other recipes.

Each recipe with a positive weight receives a number of slots in the shuffled selection order proportional to weight / min_positive_weight across all active recipes, so a recipe with weight = 2.0 is drawn approximately twice as often as one with weight = 1.0. The weight also scales the weight field on every crate::SampleTriplet returned by this recipe, which the caller’s training loop can use for loss weighting.

Recipes with weight <= 0.0 are excluded from selection entirely and no samples are produced for them.

instruction: Option<Cow<'static, str>>

Optional instruction text attached to samples from this recipe.

allow_same_anchor_positive: bool

Allow anchor and positive to carry identical text (SimCSE / dropout-trick mode).

When true, the sampler will emit triplets even when the anchor and positive sections resolve to the same text. This enables the unsupervised SimCSE training pattern: the same text string feeds both slots, and the model’s dropout layers produce two slightly different embeddings at training time.

Negatives are still required to differ from both anchor and positive.

Defaults to false; set true only for recipes whose anchor and positive selectors intentionally resolve to the same content (e.g. text-only sources).

Trait Implementations

impl Clone for TripletRecipe

fn clone(&self) -> TripletRecipe

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TripletRecipe

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for TripletRecipe

fn default() -> Self

Returns the “default value” for a type.